A comprehensive, opinionated, actionable guide for using AI to build, ship, and operate social games in the lineage covered by 🌾 The Social Games Playbook 🎮 — Stardew Valley, Township, Pixels.xyz, FarmVille 3, Dragon City, Core Keeper, etc. Read this after the main playbook. The playbook tells you what to build (the 14 pillars, the daily loop, the economy). This document tells you how to use AI to build it 5–10× faster, ship more content, and operate it intelligently — without burning yourself on legal landmines, hallucinated systems, or "AI slop" that players sniff out in 30 seconds. Distilled from current (2025–2026) tooling: Claude Code, Cursor, Unity/Godot MCP, PixelLab, Cascadeur, Inworld, Convai, Suno/Udio/ElevenLabs, ToxMod, Kumo, EA's RL playtesting, GDC 2026 sessions, Steam's January 2026 AI policy rewrite, and shipped-game case studies. If you only read three sections: §3 The Three AI Layers, §5 The 14 Use Cases (Ranked by ROI), and §17 The 90-Day Adoption Plan. 🎯 Who This Guide Is For ⚡ The 30-Second Mental Model 🧱 The Three AI Layers — Dev-Time, Ship-Time, Ops-Time 🧠 First Principles — When AI Actually Wins 🏆 The 14 Use Cases, Ranked by ROI 💻 AI for Code — The Coding Loop 🎨 AI for Visual Assets — Pixel, Sprites, UI, Concept 🕺 AI for Animation 🎵 AI for Music, SFX, and Voice 📜 AI for Narrative, Quests, Items, Lore 🗣️ Live LLM NPCs — The Danger Zone 🧬 AI Procedural Content Generation 🌐 AI for Localization 🤖 AI Playtest Bots & Economy Simulation 📊 AI for Live Ops — Churn, Segments, Personalization 🛡️ AI for Moderation — Text, Voice, Image, UGC 📣 AI for UA Creative & Marketing 💬 AI for Community & Player Support 💸 The AI Cost Stack — What an Indie Studio Actually Spends 🤝 The Hybrid Pipeline — Where Humans Stay in the Loop ⚖️ Legal, Policy, and Platform Compliance ⚠️ The Anti-Patterns — How AI Sinks Social Games 🗺️ The 90-Day AI Adoption Plan 🌱 The Greenfield AI-Native Build Plan 📋 Cheat Sheet & Tool Stack You are one of: Solo or small-team indie dev (1–5 people) building a cozy/farm/sim/sandbox game and competing with studios that have 30× your headcount. Live-ops studio operator running a Township/FarmVille-class game who needs to ship a seasonal event every 2–4 weeks without burning out the team. Web3 / crypto-native team (Pixels, Sunflower Land class) where economy balance, anti-bot, and content velocity are existential. CTO / lead at a 10–50-person studio deciding which AI bets to make in the next 6 months without committing to dead-end tooling. If you're a AAA studio with a 200-person content pipeline, this guide is still useful but the cost calculations are not your bottleneck — your bottleneck is org change. This guide assumes you have read the main 🌾 The Social Games Playbook 🎮. All references to "the daily loop," "the 14 pillars," "faucets and sinks," etc. point back there. ┌──────────────────────────────────────┐ │ AI is a force-multiplier on a │ │ CORRECT design. It does not invent │ │ the design for you. │ └──────────────────────────────────────┘ │ ┌─────────────────────────────────┼─────────────────────────────────┐ ▼ ▼ ▼ ┌──────────────────┐ ┌──────────────────────┐ ┌─────────────────────┐ │ DEV-TIME AI │ │ SHIP-TIME AI │ │ OPS-TIME AI │ │ (build faster) │ │ (in the binary) │ │ (run smarter) │ │ │ │ │ │ │ │ • Code gen │ │ • Generated assets │ │ • Churn prediction │ │ • Asset gen │ │ • Live LLM NPCs │ │ • Personalization │ │ • Playtest bots │ │ • PCG quests/loot │ │ • Moderation │ │ • Localization │ │ • Adaptive difficulty│ │ • UA creative │ │ • QA / linting │ │ │ │ • Player support │ └──────────────────┘ └──────────────────────┘ └─────────────────────┘ HIGH ROI, LOW RISK MEDIUM ROI, HIGH RISK HIGH ROI, MEDIUM RISK Use it everywhere Use it carefully Use it as you scale The single most important insight: dev-time AI compounds without risk. Ship-time AI compounds with risk (legal, quality, immersion-breaking). Ops-time AI compounds with operational complexity. Adopt in that order. Most failures come from teams doing the reverse. the binary doesn't know AI was used Tool category Examples What it replaces Risk Coding agents Claude Code, Cursor, Copilot, Windsurf Engineer hours Low Engine MCP bridges Unity-MCP, Godot AI, Unreal MCP Manual scene/asset wiring Low Asset generators PixelLab, Sprite-AI, Cascadeur, Suno, ElevenLabs Outsourcing, asset packs, junior artist Med Playtest bots RL agents, generative ABM, Chaos Dynamics Internal QA passes Low Linters / reviewers Claude review skill, security-review skill Senior eng review time Low Steam's January 2026 policy rewrite explicitly exempts dev tools (e.g., Copilot, Claude Code). They don't need disclosure. Embrace this layer fully. the binary contains AI artifacts or invokes AI at runtime Sub-layer Examples Risk Pre-generated assets AI sprite art, AI music shipped in build IP / copyright / disclosure Server-side PCG LLM-generated quest text, item names, dialogue Hallucination, drift, exploit Live LLM NPCs Inworld, Convai, on-device ACE Latency, jailbreak, cost, immersion Adaptive difficulty RL-driven enemy or pricing tuning Manipulation perception This is the layer where Steam, Apple, Google, and EU AI Act compliance live. Treat every shipped artifact as a future legal exhibit. the binary is unaware; AI runs alongside Function Examples What it replaces Churn prediction GNN models (Kumo), in-house XGBoost Guesswork on retention spend Segmentation LLM clustering of player behavior Country/level static segments Live ops orchestration AI agents scheduling events / battle pass tiers Producer hours Moderation ToxMod (voice), Hive (image), Perspective (text) Outsourced mod farms Support RAG bots over patch notes / FAQ T1 customer support tickets UA creative Sora 2, Veo 3, Higgsfield, AdCreative Video editor / motion designer hours Industry signal (2026 Unity Game Development Report): 95% of studios use AI in core workflows; 62% specifically use AI agents for backend and coding. If you don't, you're already behind on cost-per-feature. Before any tool, internalize these. exceptionally good at High-volume, low-stakes content. Crop names, item descriptions, NPC small-talk, quest variants, festival flavor text. Social games eat content like termites. Repeated structural variations. A barn, a coop, a stable, a pen — same shape, different theme. Sprite generators love this. Long-tail economy decisions. 400 items × 6 currencies × 30 levels = a balance problem humans cannot brute-force. Simulation + RL can. Behavioral pattern detection at scale. Churn signatures, bot detection, exploiters, whales-about-to-leave — classic ML wins. bad at Tone consistency across thousands of strings. AI drifts. Without a style bible and review pass, your wholesome cozy game starts sounding like a Marvel quip. Mechanical correctness. AI happily writes "you gain 5 turnips per harvest" when the spec says 3. Numbers must be schema-validated, not prose-validated. Long-arc narrative payoff. Foreshadowing across 40 hours of play. AI cannot hold this without a human story bible and tight retrieval. The "warm" feeling. Stardew Valley sold 41M copies because Eric Barone wrote every line. Players read sincerity. AI-written cozy dialogue often reads as polite-but-empty. The synthesis: use AI for volume and variation, use humans for voice, payoff, and the 100 hero strings the player remembers. Every cozy/social game has roughly 50–200 hero strings — first NPC line, marriage proposals, festival speeches, achievement unlocks, the loading-screen tip that becomes a meme. A human writes all of these. AI writes the surrounding 5,000 strings of barn-flavor and crop-tooltips. If the player would screenshot the line: human-written. Ranked for a small social-games studio (5–20 people). ROI = time saved per dollar spent, weighted for risk. # Use case ROI Risk Adopt by Notes 1 Code generation (Claude Code/Cursor) ⭐⭐⭐⭐⭐ Low Day 1 30–60% throughput gain on backend/tools. No-brainer. 2 Localization (hybrid AI+linguist) ⭐⭐⭐⭐⭐ Low Pre-launch 70–90% cost cut vs traditional LSP for first pass. 3 UA creative iteration (post-launch) ⭐⭐⭐⭐⭐ Low Soft launch TikTok needs 20–40 creatives/month; AI is the only way. 4 Pixel art / sprite generation ⭐⭐⭐⭐ Med Pre-prod Concepting: fantastic. Final assets: human polish required. 5 Churn prediction & personalization ⭐⭐⭐⭐ Med 100k MAU+ Below scale, your gut is fine. Above, GNN models pay back. 6 Voice moderation (ToxMod-class) ⭐⭐⭐⭐ Low Voice chat If you ship voice chat and skip this, you're negligent. 7 Music generation (Suno/Udio/ElevenLabs) ⭐⭐⭐⭐ Med Pre-prod Background loops great; hero theme = human composer. 8 Procedural quests / item names ⭐⭐⭐ Med Mid-prod Server-side, schema-constrained, human-reviewed. 9 Playtest bots / economy simulation ⭐⭐⭐ Low Beta Catches dead content & exploits before humans do. 10 Animation (Cascadeur, sprite-sheet AI) ⭐⭐⭐ Med Mid-prod Inbetweening + retargeting wins big; full mocap still better. 11 Player support RAG bot ⭐⭐⭐ Low Live Cuts T1 ticket volume 40–70% with patch notes + FAQ corpus. 12 Concept art & marketing key art ⭐⭐ Med Anytime Internal mood-boards: ✅. Final marketing: human-touched. 13 Live LLM NPCs (in-game runtime) ⭐⭐ High Late or never Cool demo, hard product. Read §11 before believing a vendor. 14 Voice acting (synthesis / cloning) ⭐ High Carefully Union/legal/contract minefield. Do not clone real actors. Order of adoption: start at row 1 and work down. Don't skip ahead to row 13 because it's exciting on Twitter. The single biggest lever. A solo dev with Claude Code can ship the backend a 4-person team shipped two years ago. Tool Best for Cost (May 2026) Claude Code Long-running agentic refactors, codebase-aware multi-file edits ~$20/mo Pro, $200/mo Max Cursor IDE-native pair programming, fast in-line edits $20/mo Copilot Inline completion in any IDE $10/mo Windsurf Cursor competitor, strong agent mode $15/mo Claude Code Game Studios skill pack Pre-built workflows: sprint plans, code review, asset audits, release checklists across Unity/Unreal/Godot Free, OSS Most pros run Claude Code (or Cursor) as the agent + Copilot for inline taps. Both. The latency profile is different — agents for big work, completion for typing. Model Context Protocol bridges let your AI assistant operate the engine itself: create scenes, edit prefabs, run play tests, inspect logs. Unity MCP (CoplayDev/unity-mcp) — Unity Editor exposed to Claude/Cursor. Godot AI — same idea for Godot. Unreal MCP — exists but rougher; Unreal's Blueprint serialization is a pain point. With MCP, "add a new crop type and wire it through" becomes a single conversation, not a 40-tab refactor. Set this up week 1. Add a CLAUDE.md (or .cursorrules, or AGENTS.md) at repo root. The example in this very repo at CLAUDE.md is a template. It must contain: Architecture diagram (services + data flow). Folder map (what lives where). Conventions per language (error wrapping, test style, lint config). The "common pitfalls" list specific to your repo (e.g., "never call Python service from frontend"). Build/test/lint commands the agent should run after edits. Without this, the agent invents conventions. With it, the agent is a 3-day-onboarded mid-level engineer on day 1. Use skills for repeatable workflows: /migrate, /lint, /build, /test, /review, /security-review (this repo already has them — see the available skills list). Use subagents to parallelize independent searches (e.g., "find all spawner code" + "find all loot drop code" in parallel). For balance work, never let the agent freehand numbers. Have it read a balance.yaml schema, propose changes, then run the simulation harness. Keep golden replays: deterministic save files the agent runs after every refactor to catch behavioral drift. Multi-day game-feel tuning. The AI doesn't play the game. Networking / netcode under load. It writes plausible code that breaks at p99. Shader / GPU perf optimization beyond template patterns. Anti-cheat. Adversarial reasoning needs a human security mindset. For these, AI is your typist, not your architect. Stage Tool Output Mood board Midjourney, Flux, Ideogram Style references Concept art Midjourney + ControlNet, NanoBanana Character / building concepts Pixel sprites PixelLab Game-ready sprites with 4/8 directions Sprite sheets Sprite-AI, God Mode Idle / walk / attack / hit-flash batches UI icons Recraft, Sprite-AI, custom Flux LoRA Crop icons, currency, buttons Tilesets PixelLab tileset mode, hand-tiled in Aseprite 16/32px tiles Final polish Aseprite (human) Production assets The non-negotiable: every sprite that ships gets a human pass in Aseprite. AI sprite tools in 2026 are good enough to generate, not good enough to finalize. Anti-aliasing, palette discipline, and the 1-pixel decisions that separate "indie polish" from "asset flip" still need human eyes. Players in cozy/farming Discords have an instinct for AI slop. Common giveaways: Inconsistent palette across sprites (each generation drifted). 6-fingered crop holders in NPC portraits. Tile seams that don't tile (the AI didn't understand wrap-around). Outline weight inconsistency (1px on some sprites, 2px on others). Character portrait "AI gloss" — the soft, slightly-airbrushed look from Flux/SDXL. Fix all of these in the human-polish pass. If you can't, ship fewer assets — quality > quantity in this genre, always. Once you have ~50 hand-drawn assets in the game's style, train a LoRA (on Flux or SDXL) and use it as the default generator for everything else. This is how you keep palette discipline at scale. Cost: ~$5–20 to train on Replicate/Civitai. A 32x32 pixel-art [SUBJECT], [POSE], facing [DIRECTION], [N]-color limited palette: [HEX1, HEX2, ...], 1px black outline, no anti-aliasing, transparent background, matches reference style of [GAME or LoRA name]. 4 directional variants: down, up, left, right. Iterate on the palette and pose; freeze the rest of the prompt as your house style. The main character's portrait. Players look at this 1,000 times. Pay a human. Marriage candidates' art (in dating-sim adjacent games). Same reason. Logo / wordmark. Trademark lawyers will not accept "the AI made it." Marketing key art for store listing. Steam, App Store, and Google Play all increasingly scrutinize AI key art and several have rejected listings in 2025–2026. God Mode and Sprite-AI generate idle/walk/attack/hit sprite sheets from a single base sprite. Quality: usable for prototyping; needs human cleanup for shipping. Ludo.ai sprite generator includes animation modes for indie/commercial games. Cascadeur 2026 added an AI Root Motion tool for motion style transfer — useful even for 2D devs who animate skeletal rigs. For shipping pixel animations, the realistic 2026 workflow is: AI generates the sprite-sheet skeleton (poses). Human does the inbetween cleanup and timing in Aseprite. AI is not trusted for the 8-frame walk cycle on the main character. Cascadeur — keyframe + AI physics-aware autoposing. $8/mo indie tier (commercial up to $100K revenue). Best in class for indie 3D character animation in 2026. Move.ai / DeepMotion — video-to-mocap. Replaces a mocap suit for prototyping. Rokoko + AI cleanup — same idea, more pro. AnimateDiff / runway video2anim — for cinematic and trailer work, not gameplay. Combat feel. The 4-frame hit-pause + screen-shake combo that makes Moonlighter feel good. NPC personality animations (Stardew's Pierre's hand-rub). Anything the camera lingers on. Service Quality (2026) Commercial license Best use Suno v5 Excellent Unsettled. Settled with WMG; Sony lawsuit pending summer 2026 Demo / prototype / temp tracks Udio Excellent Settled with UMG; UMG-Udio joint platform launching 2026 Track generation; pivot when joint platform launches ElevenLabs Music Good Clean. License-clean enterprise terms Shippable background tracks Stable Audio Good (loops) Clean (Stability commercial) Loopable ambient / sting beds Riffusion OK (loops) Clean Ambient / variation AIVA Good Clean (Pro tier) Orchestral / cinematic Practical rule for shipped music in 2026: use ElevenLabs Music, Stable Audio, or AIVA Pro. Use Suno/Udio for prototype and trailer scratch only until their licensing fully settles. If your game ships a Suno track and Sony wins its case, you have a takedown problem. The Business Tycoon case study is the proof point: 4× 2-minute instrumental tracks, ~2 minutes total generation time, $3.20. That's the new floor for background-music cost. The main menu theme and the song that plays when the player gets married / completes the museum / wins the festival is human-composed. Always. This is your "Stardew Valley Overture." Players associate it with the brand for a decade. Outsource it: $500–3,000 from a Fiverr Pro / Soundcloud composer or $5–20K from a name like ConcernedApe-tier indies. Don't generate it. ElevenLabs Sound Effects — text-to-SFX, license-clean. Ship-ready. Adobe Audition + AI denoise / cleanup — for human-recorded foley. Soundly / Splice — non-AI but deserves a slot in the stack. For a farming/cozy game you need ~200 SFX (tool swings, UI clicks, ambient layers, footsteps × surface, animal sounds). Generating with ElevenLabs: ~$30 in credits, ~1 day of curation. This is the highest-risk AI sub-domain. Use case Recommendation Full VO for cozy NPCs Skip — most cozy games have no VO; preserve the player's inner reading voice. Short barks / greetings ElevenLabs voices, original / synthetic, never cloned. Narrator Hire a human (it's 50–200 lines, the most player-facing audio in your game). Cloning a real actor Don't. Even with consent, US/EU contract law, SAG-AFTRA agreements, and likeness rights make this a multi-year liability. Live LLM NPC voice (§11) If you ship this, pre-license cloned voices via Inworld/ElevenLabs Enterprise with full contract chain. This is where AI most reliably 10×s your throughput in social games — if you constrain it properly. Never let an LLM emit free-form game content. Always emit structured JSON validated against a schema. Example: { "id": "quest_spring_radish_001", "giver_npc": "pierre", "season": "spring", "tier": 1, "title": "<= 40 chars, no emoji, sentence case", "description": "<= 220 chars, second person, cozy tone", "objective": { "kind": "deliver", "item": "radish", "qty": 5 }, "reward": { "gold": 120, "xp": 30, "friendship": { "pierre": 1 } }, "tone_tags": ["wholesome", "low_stakes"] } The LLM fills the fields. A schema validator (Zod, Pydantic, JSON Schema) rejects malformed output. A balance validator rejects rewards outside the curve in your balance.yaml. A tone-checker LLM does a second pass to flag off-voice strings. This pattern alone is the difference between "AI quest generator that ships" and "AI quest generator that floods QA with garbage." For a Township-class game, AI should generate: 200–500 collection quests (deliver X to Y). 100–300 item descriptions. 50–200 NPC small-talk lines per character (5 characters = 250–1000 lines). 30–60 festival flavor strings per festival. 50–100 loading-screen tips. Crop / animal / building names and 1-line descriptions. Hero strings (still human): NPC introductions, romance arcs, festival speeches, achievement unlocks, the endgame letter, the player's wedding. A 2–4-page document the LLM reads on every generation request: Tone words (e.g., "warm, gently witty, never sarcastic, never edgy"). Tone anti-words ("avoid: cynical, ironic, modern slang, references to social media, profanity"). Voice samples per NPC (3–5 lines of hand-written dialogue each). Forbidden topics (politics, real-world religion, modern tech). Punctuation and capitalization rules. Example accept / reject pairs. Without this, every generation drifts toward GPT-default voice (which is the voice of a polite-but-bland LinkedIn post). Model Best for Notes Claude Opus 4.7 / Sonnet 4.6 Long-form narrative, tone-sensitive prose Best tone fidelity; the default GPT-5 / GPT-5-Pro Structured JSON-mode generation, fast bulk Fastest with json_schema Gemini 2.x Pro Long-context lore consistency (1M+ ctx) Good when feeding the whole story bible Open-source (Llama, Qwen) Offline / cost-floor / uncensored variants Self-host; useful at very high volume Always cache. Your style bible is reused on every call. Anthropic / OpenAI / Gemini all support prompt caching — it cuts cost 50–90% for static system prompts. A typical content-gen pipeline pays $0.0001–0.001 per generated quest after caching. The shiny demo. The hardest production system. Read this whole section before deciding. Inworld AI — Character Engine; powered the GDC 2024 Covert Protocol demo (NVIDIA + Inworld), now used in a handful of indie titles and VR games (Office Whispers, etc.). Convai — LLM NPCs with the Actions feature (LLMs trigger in-game actions, not just dialogue). NVIDIA ACE — runs on-device on RTX hardware as of 2026; removes the cloud roundtrip. Open-source (AkshitIreddy/Interactive-LLM-Powered-NPCs et al) — works for solo devs, not production-hardened. Social games are about persistence, predictability, and the warmth of recognition. "Pierre says the same thing on Wednesday" is a feature. Players come back because their world is comfortingly stable. An LLM NPC is the opposite: stochastic, novel, sometimes inconsistent. This is great for an immersive sim or detective game (Covert Protocol), and culturally wrong for a Stardew-class cozy game. Players will ask Pierre about Bitcoin, Pierre will answer, the immersion breaks. [ ] Personality + memory persisted server-side, never trusted from client. [ ] Hard knowledge boundary: NPC knows their lore, refuses out-of-world topics in-character ("I don't know what 'Bitcoin' is, friend"). [ ] Topic blocklist for politics, real-world tragedies, sexual content, self-harm. [ ] Latency budget under 1.5s for first audio token (otherwise dialogue feels broken). On-device ACE or streaming TTS required. [ ] Cost budget: $0.001–0.01 per turn × millions of turns. Model this before committing. [ ] Jailbreak red-team before launch; reproduce attempts post-launch via telemetry. [ ] Disclosure on Steam/App Store per January 2026 policies. [ ] Fallback to scripted dialogue if the LLM service is down. [ ] Per-player rate limits to prevent abuse / cost runaway. [ ] Voice cloning contract chain if the NPC has a voice (do not skip — see §9.4). Instead of full LLM NPCs, use LLMs at design time to write 10× more scripted dialogue, then ship that scripted dialogue. Players get the feel of a fuller world without runtime risk. This is what most successful cozy games will do for the next 3–5 years. If you must ship runtime LLM behavior, scope it tight: LLM controls only side characters (a wandering bard, a stranger at the inn). Core characters (marriage candidates, family, vendors) stay scripted. LLM output is constrained to a topic whitelist ("the inn, the weather, local rumors"). Live AI-generated content must be disclosed on the store page. Live AI-generated adult sexual content is an absolute prohibition with no exception — relevant if your social game has romance and you let a runtime LLM handle it. Don't. Apple and Google have parallel policies; expect tightening through 2026. System PCG fit Notes Daily orders / quests Excellent Bounded, schema-driven, low narrative weight Item / crop / animal names Excellent Pure flavor; cap collisions with a uniqueness check Dungeon / mine layouts Good Wave Function Collapse + LLM hints for set dressing World / island generation Good Minecraft-class; deterministic seed + LLM biome flavor Loot drops Good Constrained generation against an item DB NPC names + 1-line bios Good For populating festivals, leaderboards Main story arc Bad Players need authored emotional payoff Romance dialogue Bad Same Tutorial Bad Must be deterministically correct [Player request / time tick] │ ▼ [Server PCG service] │ ├─► Fetch context (player level, inventory, season, last 7 days of quests) │ ├─► Build prompt with style bible + schema │ ├─► LLM generate (with prompt cache) │ ├─► Schema validate ──► reject + retry on fail │ ├─► Balance validate ──► clamp values to curve │ ├─► Tone validate (cheap second LLM pass) ──► flag for human │ ├─► Persist to DB │ └─► Return to client Never call the LLM from the client. Every generation runs on your server, with rate limits, caching, and validation. This also gives you the audit log you'll need under EU AI Act requirements. Set temperature low (0.2–0.5) for items / quests where players will compare in Discord ("did you get the carrot quest? me too"). Set higher (0.7–0.9) for personal flavor strings (loading-screen tips, idle barks). Use a seed derived from player ID + day so the same player gets the same daily content even on retry. This prevents save-scumming and fairness complaints. Maybe the highest-ROI use case after coding. Traditional LSPs charge $0.10–0.20 per word. AI-first hybrid pipelines charge $0.01–0.03 per word at equivalent quality for a cozy/casual game. Source strings (en) │ ├─► Translation Memory match (free) [exact / fuzzy reuse] │ ├─► AI MT first pass (Claude / GPT / DeepL Pro) [bulk volume, $] │ └─ with: glossary, style guide, character voice notes, screenshots │ ├─► AI tone/cultural review (second LLM pass) [flags for human] │ ├─► Human linguist review [transcreation, hero strings] │ └─► QA pass in-game (LLM screenshot review) [overflow, truncation, missing vars] Alocai — game-specific MT + GenAI (ModelWiz). Gridly — string management with AI translation built-in. Lokalise + AI — established LSP platform, now AI-augmented. Custom Claude/GPT pipeline — for studios with engineering capacity; offers most control. out of the box Spanish, Portuguese (BR), French, German, Italian, Polish, Russian, Korean, Japanese, Simplified Chinese. Japanese — honorifics + character voice = automated MT will break tone in cozy games. The MT first pass is fine; the linguist pass is mandatory. Korean — same. Arabic — RTL layout, dialect variation, cultural sensitivities (alcohol, religion). Traditional Chinese — different from Simplified in tone and idiom; treat as separate. Thai / Vietnamese — tonal nuances and segmentation issues. AI lip-sync + voice cloning makes 10+ language full VO feasible for indie budgets in 2026. For a cozy game with no VO, don't add VO just because you can. For a game that has VO, AI dubbing of side characters is acceptable; main cast = human VO per language as far as budget allows. Build a glossary table on day 1: EN term Tone ja-JP ko-KR de-DE Notes Energy warm げんき 활력 Energie Not "stamina" Coin (currency) neutral コイン 코인 Münze Singular always Mayor warm 村長 촌장 Bürgermeister Honorific in jp/kr This glossary feeds into every AI translation call. Without it, "Energy" becomes 5 different words across your game in the same language. EA's RL-driven playtest framework (publicly described in 2024–2025) caught: Inconsistent AI behavior at edge cases. Balance asymmetries between teams. Physics / animation glitches. Unreachable content. Stuck states that human QA never reproduced. For a social game, the equivalent is: Economy traps — quests that lock the player out of progression. Dead content — items no rational agent ever buys. Exploit routes — recipes / arbitrage loops that print money. Difficulty walls — levels where the optimal strategy still fails 80% of the time. Energy starvation — sequences where the player runs out of energy before the next milestone. Build (or buy) an agent-based simulator that replays your economy with thousands of synthetic players, each with a different strategy: "Greedy gold-maximizer" "Completionist" "Casual 2-sessions-a-day" "Whale spender" "F2P optimizer" "Bot operator" Run it before every economy patch. Outputs: Currency inflation curves. Gini coefficient on wealth across cohorts. Time-to-paywall by archetype. "Dead recipe" report. Exploit yield (gold-per-hour for the optimal exploit found). For LLM-based realism, recent research (arXiv 2506.04699 / 2512.02358) demonstrates Generative Agent-Based Modeling — LLMs fine-tuned on real player logs play your game and surface emergent behaviors traditional ABM misses. Worth the investment at MMO scale; overkill for prototypes. Roll your own. A 500-line Python harness running 10K simulated players overnight catches 80% of economy bugs. Highest ROI per engineer-week. Chaos Dynamics — commercial high-fidelity simulation. Unity ML-Agents — for engine-integrated RL playtesting. OpenAI / Anthropic LLM agents orchestrated via tool-use to play the game over a real network. Pull from the main playbook §20 (KPIs). The simulator should output all of them for every release candidate. If you can't simulate them, you can't iterate fast enough to compete. Live ops is the multi-year game in social-games. AI here pays back over years. Stage Approach < 10K MAU Don't bother. Your gut + cohort tables are enough. 10K–100K MAU XGBoost / LightGBM on session + monetization features. Internal data scientist can build in 2–4 weeks. 100K–1M MAU XGBoost still wins; add survival models for time-to-churn. 1M+ MAU Graph Neural Networks (Kumo, in-house PyG). Friend-graph signal is the differentiator. The Kumo case study figure: 5M MAU × 20% monthly churn among monetizers can yield ~$18M/year savings from a 10% retention lift on at-risk spenders. The math at smaller scales is proportional. Personalization layer What's safe What crosses the line Difficulty (PvE only) Slight enemy HP / spawn-rate tuning to keep flow Hidden difficulty adjustment that punishes wins Daily quest selection Bias toward content the player engages with Hiding content the player would enjoy Push notification timing Send when player historically opens Manipulative urgency / fake-scarcity FOMO Offer composition Bundle items the player has searched for Hidden price discrimination (illegal in EU) Friend / guild suggestions Match by play-time overlap and level Sorting by predicted spend EU Digital Services Act + AI Act + consumer protection law actively police this. Personalize for engagement and joy, not exploitation. The Civil War of 2025–2026 lawsuits against gacha / loot box mechanics is a preview. A single Claude/GPT agent, run on a daily cron, with read-only access to your analytics warehouse, can: Diagnose why DAU dropped 4% yesterday. Suggest which event slot to fill next based on cohort fatigue. Draft a battle-pass tier list and write the patch notes. Flag anomalies: "Crop X consumption is 20σ above baseline — check for exploit." Generate an exec summary email by 9am. Build this. It replaces 10 hours of producer work per week. Web3 and F2P social games attract botters. ML signals: Inhuman session regularity (variance below human noise floor). Click pattern uniformity. Wallet clustering (Web3). Cohort sharing (multi-account farm). Graph centrality in the trade network. GNNs win again here. Off-the-shelf: Sift, Kasada, DataDome. In-house if Web3. If your social game has chat, voice, UGC, or trade — you need moderation infrastructure on day 1. Skipping this is the #1 mistake of Web3 games and live-ops games alike. Surface Tool Coverage Text chat Perspective API, OpenAI / Anthropic moderation, custom LLM filter Slurs, harassment, grooming, spam Voice chat ToxMod (Modulate) Real-time toxic-voice detection, integrates with Discord SDK as of Jan 2026 Image / UGC Hive Moderation, Sightengine NSFW, violence, hate symbols Player names Custom blocklist + LLM check Slur variants, trademark abuse Trade / market Pattern detection + LLM intent check Scam detection, real-money trade Forums / Discord AutoMod + custom LLM workflows Brigading, off-topic, doxxing The Call of Duty case study is the public proof: 50% reduction in toxicity exposure (CoD MWII multiplayer + Warzone NA). 25% reduction in toxicity exposure (CoD MWIII global ex-Asia). 8% month-over-month reduction in repeat offenders. For a social game with voice (rare in cozy, common in MMO/sandbox), this is the only currently mature voice moderation product. As of January 2026 it integrates with Discord's Social SDK, which is how a lot of indie games already handle voice. Signal → Auto-action (mute, shadow-ban, throttle) → Human moderator queue → Player appeal → Audit log Never auto-ban without an appeal path. Never train your model on appeals you didn't review. Keep the audit log for 90+ days for both legal and false-positive review. Post-launch, your survival depends on creative velocity. This is the lever AI was built for. TikTok generated $28B in 2025 ad revenue; for mobile games, it is now often cheaper CPI than Meta but creative-heavy. TikTok algorithm rewards creative velocity: 7–10 day fatigue window vs Meta's 2–3 weeks. Minimum viable cadence for a serious mobile UA program: 20–40 creatives/month per major channel. A 4-person UA team cannot manually edit that. AI is the only way. Tool / Model Output Use for Sora 2 Photoreal video, 10–30s UGC-style testimonials, gameplay-cuts Veo 3 Video, strong physics Same Runway / Kling Video generation, image-to-video Stylized cuts Higgsfield Ads Game screenshot → ad video in 3 clicks Programmatic creative variations AdCreative.ai Static + variants Static placements, banner sets ElevenLabs Voice-over for ads Multi-language ad VO Claude / GPT Hooks, taglines, ad scripts Pre-production ideation Segwise / your MMP Performance feedback loop What's winning, what's fatigued Brief → AI variant gen (50–200 variants) → Cheap broad test ($300–1000) → Top 5% scaled → Performance feedback → New brief based on winning hooks The studios winning UA in 2026 are running this loop weekly per channel. If you're shipping 4 creatives a month, you're getting outbid. The launch trailer. Your one piece of art that lives forever on YouTube and your store page. Hire a game-trailer studio. Festival / Steam Next Fest creative. Higher-stakes attention; humans matter. Community-fan content. The single most credible creative is a streamer playing your game. The hook concept itself. AI can produce 200 variants of a hook; it rarely invents the new hook. Humans set direction; AI executes the variations. Build it on day 1 of soft launch. Inputs: Patch notes (ingested daily). FAQ (curated weekly). Game wiki / lore (slow-changing). Common ticket categories with canned answers. Output: a Discord bot + in-game help widget that handles 40–70% of T1 tickets. Common stack: Claude/GPT + a vector store (Pinecone, Weaviate, Postgres pgvector) + a thin web service. Player message → RAG bot answer → "Did this help?" → If no, route to human queue → Human answer → fed back into FAQ Two non-negotiable rules: The bot must be allowed to say "I don't know — connecting you to a human." Hallucinated answers about refunds and account issues are how you end up in a regulator's inbox. Human responses become future training data. Build the loop. Run an LLM agent daily across: Steam reviews (delta vs last week). Discord top channels (digest). Reddit subreddit (top posts + sentiment). App Store / Google Play reviews. Twitter/X mentions. Output a 1-page exec summary: top 3 complaints, top 3 praises, notable streamer/influencer activity, sentiment delta. Replace the producer's manual community scan. Cost: $5–20/day in API spend. Realistic monthly spend for a 5-person social-games studio in 2026 (USD): Layer Service Monthly cost Coding agents (per dev) Claude Code Max + Cursor + Copilot $100–250 Asset generation PixelLab + Cascadeur Indie + Flux $30–80 Music + SFX ElevenLabs + AIVA Pro $30–80 Localization (per release) AI MT + linguist (10 langs, ~5K w) $200–600 LLM content generation Anthropic / OpenAI API + caching $50–500 Playtest simulation compute AWS / GCP spot (overnight runs) $50–200 Live LLM NPCs (if applicable) Inworld / Convai Pro $200–2000+ Voice moderation ToxMod (per concurrent voice user) scaled Text moderation Perspective / OpenAI mod (free–$) $0–100 UA creative generation Sora 2 + Higgsfield + Runway $200–1000 Analytics LLM agent Claude / GPT API $50–200 Total for a pre-launch indie team: ~$700–1,500/month. For a live-ops studio doing serious UA: $3,000–10,000/month. Compare to: One outsourced pixel artist: $2–5K/month. One translator across 10 languages, traditional LSP: $5–15K/release. One UA creative agency: $5–20K/month + media. One T1 support agent: $3–6K/month. The math has been favorable since mid-2024 and the gap has widened every quarter since. Track per-feature cost. After 3 months you'll find: 60–70% of LLM spend is on a single workflow (usually content gen or live-ops agent). Caching cuts that 50–80%. Open-source models (Llama, Qwen, DeepSeek) handle 30–60% of low-stakes calls at 10× cheaper. Tier your model usage: cheap model for first pass, expensive model for hero strings, frontier model only for narrative-critical generations. The summary table for "what does AI do, what does a human do" across the pipeline: Function AI does Human does Code Bulk, refactors, tests, boilerplate Architecture, netcode, anti-cheat, perf Concept art Mood boards, 100 variations Final direction, hero key art Pixel sprites Generation, sprite-sheet expansion Final polish in Aseprite, hero portraits Animation Inbetweening, retargeting, sheet expansion Combat feel, NPC personality, camera frames Music Background loops, ambient beds Hero theme, festival music, brand jingles SFX 90% of library Signature sounds (level up, harvest) VO Side characters (if any) Main cast, narrator Quest text Bulk variants, tooltips, item descriptions Hero strings, romance arcs, story beats Localization First pass MT, glossary, cultural flag Hero string transcreation, JP/KR/AR review QA Smoke tests, regression, exploit hunting Game-feel QA, "vibes" QA Live ops Anomaly detection, churn prediction, draft patch notes Final calls on events, balance, comms UA creative Variant generation, copy variants Brief, brand voice, launch trailer Support T1 RAG, sentiment digest T2/T3, refunds, escalations, comms Moderation Detection, triage, auto-action Appeals, novel cases, policy updates Playtest RL bot exploration, balance simulation Game-feel playtests, "is this fun" calls Read across: AI handles 60–80% of the volume in every row. Humans own the 20–40% that defines whether the game has a soul. Dev tools (Copilot, Claude Code, Cursor) — exempt; no disclosure required. Pre-generated assets shipping in the build — disclosure required on store page (AI generation kind, content types). Live AI generation at runtime — disclosure required, plus you certify guardrails. Live AI-generated adult / sexual content — prohibited, no exception. Failure to disclose → store removal risk. Increasing scrutiny on AI-generated key art and screenshots. Apps with live LLM features must have content moderation pipelines disclosed. App Review will reject games that allow uncontrolled LLM output, especially for under-13 ratings. Several documented rejections in 2025 of games that didn't disclose AI-generated marketing assets. Similar disclosure expectations as Apple. Active enforcement on deepfake / impersonation / explicit AI content. Targeted ad / personalization disclosures aligning with EU norms. Most social games will fall under "limited risk" (transparency obligations): Inform players when interacting with an AI system (live LLM NPCs, AI moderation). Label AI-generated content where reasonable. Higher-risk if you do AI-driven personalization that materially affects player welfare or finances. US Copyright Office: works without meaningful human creative input are not protected. Translation: "I prompted Midjourney for the box art" likely cannot be copyrighted. "I prompted, then a human extensively edited, layered, composited, and directed" likely can. Train model warranties: get indemnification from your AI provider against third-party IP claims — Anthropic, OpenAI, Google, ElevenLabs, Adobe Firefly all offer some form of this for enterprise tiers. Free / consumer tiers usually do not. Cloning a real person's voice without consent is actionable in most jurisdictions and explicitly prohibited by SAG-AFTRA agreements. Even with consent, get a written, signed, scope-limited license. "Use my voice for game X for 5 years in markets Y, in genre Z, with the option to extend at price W." Synthetic voices with no human clone source are lower-risk but still need provider warranty. Don't train your customer-service models on player chat without a consent path. Don't feed player payment / PII data into 3rd-party LLM APIs without DPA in place. Anthropic / OpenAI / Google enterprise tiers all have zero-retention modes — use them for any pipeline touching player data. These are the failures we see repeatedly. Avoid each. It won't. AI does not know whether your daily loop is satisfying. AI does not playtest your economy on a real Wednesday with a real distracted player. Use AI to implement your design, not invent it. Players in cozy/farming Discords will identify AI sprites in 30 seconds and broadcast it. The marginal cost saved on assets is dwarfed by the wishlist hit you take in week 1. Either polish AI assets to invisibility or commission human work. A demo of a chatty NPC is not a feature. It's the easy part of a system that must include: persona persistence, jailbreak defense, cost control, latency budgets, content moderation, fallback paths, and disclosure. Most teams underestimate this by 5–10× engineering weeks. See §11. Without a 2–4 page style bible, every LLM call drifts toward the same flat "GPT-cozy" voice. By string #500 your game sounds like a content farm. Write the style bible first. Numbers go in balance.yaml. Strings go in strings.json validated by schema. The LLM never invents quantities. Every shipped data point passes a validator. Skip this and you'll ship "Deliver -1 carrots for ∞ gold" within 2 weeks. Anthropic, OpenAI, Google all have outages and price changes. Build a model-abstraction layer (or use one — LiteLLM, OpenRouter, your own thin wrapper) so you can swap. Especially important for live-runtime systems. Risk profile: a Sony win in summer 2026 could force takedowns of trained content. Use license-clean alternatives (ElevenLabs Music, Stable Audio, Adobe Firefly Audio, AIVA Pro) for anything in the build. Use Suno/Udio for trailers, scratch, and prototypes only. Dynamic difficulty that makes the player lose more right before an offer. Hidden price discrimination. Fake-scarcity push notifications. These are illegal in EU consumer law and shameful regardless. Personalize for delight, never for extraction. It is January 2026. Steam, Apple, Google, and EU all have disclosure regimes. The cost of disclosure is a paragraph on a store page. The cost of non-disclosure is store removal. Disclose. Auto-ban systems with no appeal path will produce a 1–5% false-positive rate, which at 100K MAU = 1,000–5,000 wrongly banned players per month. Each one is a refund, a chargeback, a Reddit thread, a review-bomb. Always have a human appeal path. The team sizes work because the senior person knows what AI is doing wrong. Replacing your only senior with juniors-plus-Claude is how you ship a game that's half-built and unfixable. Start with senior + AI; add juniors later. Don't claim "hand-crafted by humans" on Steam if your sprites are AI. Don't pretend your live NPCs are pre-scripted. Players will find out. Communities are forensic. The trust damage outweighs anything you saved. For an existing 5–20 person social-games studio not yet AI-native. [ ] Every developer on Claude Code (or Cursor) + Copilot. Standardize. [ ] Repo-root CLAUDE.md / .cursorrules written. (Use this repo's CLAUDE.md as a template.) [ ] Unity-MCP / Godot AI installed; one engineer demos a scene-edit conversation in standup. [ ] Style bible drafted (2–4 pages). [ ] Glossary spreadsheet started. [ ] One "champion" appointed per discipline (code, art, audio, narrative, ops). [ ] Schema-validated content generation pipeline live for items + quests. [ ] AI translation pipeline for one new language end-to-end (pick the cheapest: Spanish or Portuguese). [ ] Pixel-art LoRA trained on existing house style. [ ] AI playtest harness scaffolded; runs nightly. [ ] RAG support bot built on patch notes + FAQ (internal-only first). [ ] First content pack shipped with AI-generated bulk content + human hero strings. [ ] Localization to 3 languages shipped via hybrid pipeline. [ ] UA creative iteration loop running on TikTok/Meta — 20+ creatives/month minimum. [ ] Live-ops agent producing daily exec summaries. [ ] Moderation stack (text minimum; voice if applicable). [ ] Disclosure language updated on store pages. [ ] Churn prediction model live (if MAU justifies). [ ] AI-generated asset pipeline integrated into sprint cadence. [ ] Cost dashboard per-feature; tier models (cheap for bulk, frontier for hero). [ ] Postmortem: which AI bets paid, which didn't. Cut what's underperforming. [ ] Hiring plan adjusted: which roles do you still need, which do you not, which new ones (data scientist? RL eng?) do you? You are now operating at ~2× the throughput of a non-AI peer studio at ~70% of the cost. You will get outpaced by competitors who started 6 months earlier. Keep iterating; don't celebrate. For a brand-new social game starting fresh in 2026. AI for mood boards, references, prototype mock-ups. Cheap, fast, throwaway. AI for competitor analysis — feed AppMagic / SensorTower exports + Steam reviews into Claude/GPT, ask for tonal differentiators. A human writes the design pillars. AI does not. One engineer + Claude Code + Unity-MCP / Godot AI builds the daily-loop prototype. AI generates the placeholder art at full volume; the artist polishes the 50 hero assets. Human composer writes the hero theme; AI fills the 8–12 background loops. All numbers in balance.yaml. All strings in strings.json. Schema-validated. From day 1. Schema-driven LLM content gen for 200+ quests, 300+ items, 500+ NPC barks. Style bible enforced on every gen call. LoRA trained; sprite pipeline runs at 10× original throughput. AI playtest bots running nightly; balance issues caught before human QA sees them. 3 launch languages via AI hybrid pipeline. UA creative iteration loop spinning at 30+ creatives/month per channel. Moderation stack live before any voice/chat opens. RAG support bot live; CS agent supervising it. Live-ops agent running daily exec brief. Disclosure language reviewed by counsel and live on the store page. Full localization (10+ languages). Churn prediction online. Personalization layer running — engagement-positive only, regulator-compliant. Full live-ops cadence: 2–4 week event drumbeat, AI doing 60–80% of content, humans owning the 20% players remember. The thesis: a 4–6 person team can ship and operate, end-to-end, what a 25-person team shipped in 2022. Layer Pick Backup option Coding agent Claude Code (Max tier) Cursor Inline coding GitHub Copilot Codeium Engine bridge Unity-MCP / Godot AI Custom MCP server Concept art Midjourney v7 / Flux Pro Ideogram Pixel sprites PixelLab Sprite-AI Sprite animation Sprite-AI / God Mode Manual Aseprite 3D animation Cascadeur Indie Move.ai Music (shippable) ElevenLabs Music + AIVA Pro Stable Audio SFX ElevenLabs Sound Effects Splice / Soundly Voice synthesis ElevenLabs (synthetic only) OpenAI TTS LLM content gen Claude Sonnet 4.6 + Haiku 4.5 (tiered) GPT-5-Pro / GPT-5 Live LLM NPCs (if shipping) Inworld AI Convai Localization Custom Claude pipeline + linguist Alocai / Gridly Playtest bots Custom Python + Unity ML-Agents Chaos Dynamics Churn ML XGBoost (in-house) / Kumo LightGBM Voice moderation ToxMod (no real competitor in 2026) Text moderation OpenAI moderation + Perspective Custom LLM filter Image moderation Hive Moderation Sightengine UA creative video Sora 2 / Veo 3 + Higgsfield Ads Runway Player support Custom RAG (Claude + Postgres pgvector) Intercom Fin Analytics agent Claude / GPT scheduled cron Hex / Mode + LLM extension When deciding whether to add AI to a workflow, ask in order: Is the input bounded by a schema? If yes → AI is safe. If no → wrap it. Is the output reviewable in <30 seconds by a human? If yes → ship it. If no → automate the review. Is the failure mode embarrassing or expensive? If yes → human in the loop. If no → trust automation. Is the task high-volume, low-stakes? Perfect AI fit. Is the task low-volume, high-stakes? Keep it human. Does a regulator care about this output? Disclose, log, audit. Would the player screenshot this? Human owns it. Install Claude Code / Cursor + Copilot for every dev. Install Unity-MCP or Godot AI in your engine. Write a 2-page style bible. Move all numbers to balance.yaml, all strings to strings.json. Set up a schema-validated content-gen prototype on one quest type. Pick one language (Spanish) and run the AI hybrid localization end-to-end on 200 strings. Build the daily live-ops AI agent and pipe its output to your team Slack at 9am. You will measurably ship faster within 2 weeks. Compounding starts immediately. AI scales the parts of social games that don't have a soul, so humans can spend their time on the parts that do. If you keep that line in mind on every adoption decision, you'll get most of these calls right. The companion to this document: 🌾 The Social Games Playbook 🎮 — the design playbook this AI guide is built to accelerate. Steam AI policy (Jan 2026): https://store.steampowered.com (Valve disclosure requirements) 2026 Unity Game Development Report — AI adoption stats. GDC 2026 AI in Game Development track — recordings via the GDC Vault. arXiv 2410.15644 — PCG in Games: Survey with Insights on LLM Integration. arXiv 2506.04699 — Generative Agent-Based Modeling for MMO Economies. arXiv 2512.02358 — Beyond Playtesting: Multi-Agent Simulation for MMOs. Modulate / ToxMod case studies (Activision, Schell Games). Anthropic / OpenAI / Google enterprise data-use and indemnification terms. This document is a living guide. AI tooling moves quickly — re-evaluate every 90 days. The principles in §3, §4, and §22 should outlast the specific tools. If you found this helpful, let me know by leaving a 👍 or a comment!, or if you think this post could help someone, feel free to share it! Thank you very much! 😃